Meta Andromeda: Powering the Next Generation of Ads Retrieval Through Model-System Co-Design

Why it matters

Meta has unveiled the next generation of Meta Andromeda, its machine learning system for personalized ads retrieval, developed in close collaboration with NVIDIA. Retrieval is the opening stage of the ads recommendation pipeline, narrowing an enormous pool of eligible ads to a small set of candidates within milliseconds before a costlier ranking stage scores them in detail. Historically, the deep modeling quality of ranking was too expensive to apply at retrieval scale, but this redesign co-optimizes the model, system, and GPU hardware to bring near ranking-level performance to the retrieval stage. The shift was driven by practical pressure: generative AI creative tools have flooded the platform with vast numbers of new ads, forcing retrieval to evaluate a far larger candidate pool within the same latency budget. The new architecture borrows mixture-of-experts ideas from large language models, using specialized user towers, context-aware ad encoding, and attention-based gating to route candidates efficiently. Meta reports the system has already lifted conversions on Facebook since its rollout began.